[PATCH 3/3] sched_ext: idle: Reuse sd_llc_shared->has_idle_cores

[Andrea Righi <arighi@nvidia.com>]

Rely on sd_llc_shared->has_idle_cores to determine if we can skip the
scan for a fully idle SMT core in the same LLC.

This can save some expensive cpumask scans by quickly checking the
cached LLC idle cores information.

At system saturation, this change appears to reduce scheduling latency
for interactive workloads, resulting in a consistent 2–3% fps speedup
for schedulers using the built-in idle CPU selection policy.

Tested with scx_cosmos, scx_bpfland and scx_rustland.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 kernel/sched/ext_idle.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/ext_idle.c b/kernel/sched/ext_idle.c
index d2434c954848e..92a9a7edc2b17 100644
--- a/kernel/sched/ext_idle.c
+++ b/kernel/sched/ext_idle.c
@@ -558,6 +558,8 @@ s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
 	 * partially idle @prev_cpu.
 	 */
 	if (sched_smt_active()) {
+		bool has_idle_cores = test_idle_cores(prev_cpu);
+
 		/*
 		 * Keep using @prev_cpu if it's part of a fully idle core.
 		 */
@@ -570,12 +572,17 @@ s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
 
 		/*
 		 * Search for any fully idle core in the same LLC domain.
+		 *
+		 * Skip this step if we already known this LLC has no fully
+		 * idle cores.
 		 */
-		if (llc_cpus) {
+		if (llc_cpus && has_idle_cores) {
 			cpu = pick_idle_cpu_in_node(llc_cpus, node, SCX_PICK_IDLE_CORE);
 			if (cpu >= 0)
 				goto out_unlock;
 		}
+		if (has_idle_cores)
+			set_idle_cores(prev_cpu, false);
 
 		/*
 		 * Search for any fully idle core in the same NUMA node.
-- 
2.51.1